Generalization of English-only findings about compression and performance

Ascertain the extent to which findings from English-only experiments on the relationship between tokenization compression and language model performance generalize to other languages, including typologically diverse and low-resource languages.

Background

Some controlled studies limited to English report no relationship between compression and performance, raising the question of whether such results hold for other languages with different scripts, morphology, and tokenization behavior.

Establishing crosslingual generalization is vital for equitable multilingual model development and for understanding whether tokenizer design choices should differ by language.

References

These experiments are limited to English, and it is thus unclear the extent to which they generalize to other languages.

— Explaining and Mitigating Crosslingual Tokenizer Inequities  (2510.21909 - Arnett et al., 24 Oct 2025) in Section 2.3 (Related Work — The Relationship Between Compression and Performance)

How our findings transfer to low-resource languages, and to scripts with markedly different orthographic conventions, remains an open question that we leave to future work.

— Objective vs. Search: Decomposing What Makes a Good Tokeniser  (2609.19145 - Yavuz et al., 16 Sep 2026) in Limitations, Languages paragraph

The ablation uses the smaller E2B at a fifth of the final budget, and stronger initializers \citep{wechsel2022,focus2023} remain untested, so extension could become viable at larger scale.

— From Zero to Hero: An Open LLM Ecosystem for Armenian  (2609.03350 - Arakelyan et al., 3 Sep 2026) in Section ‘Adapting Gemma-4 to Armenian’, paragraph ‘Keep or extend the tokenizer?’

One mechanism that we hypothesize, which our experiments do not isolate, is overfitting of the token embeddings: the inverted-allocation tokenizers segment the low-resource languages into fewer, longer, more language-specific vocabulary entries (Tamil's fertility falls to $0.379$ times the reference tokenizer's; \cref{app:inversion}). These tokens necessarily occur in a smaller set of training contexts, and repeating those contexts may overfit the embeddings to them.

— Language-Specific Effects of Tokenizer Choice in Multilingual Language Models  (2610.12144 - Meister et al., 8 Oct 2026) in Appendix, Section “Equal Weight and Resource-inversion Experiments,” subsection “Exclusion and reallocation compared,” paragraph beginning “One mechanism that we hypothesize.”

We therefore leave the cause of the saturation difference unresolved.

— Neurograf v1: An End-to-End Romanian-First Language Model Family  (2610.13134 - Rotaru, 9 Oct 2026) in Section 3.6